jetson-arena

How does a given Jetson device handle a given model setup?

A setup can be a single model, a mixture of models, or a whole pipeline under a pinned config. The flagship example is still the realtime VAD → STT → LLM → TTS loop: four models on one board, contending for VRAM, memory bandwidth, and the GPU — each fits alone; the setup is what fails. jetson-arena benchmarks setups for latency, memory signature, and quality, then publishes the results together with the Docker build recipes that reproduce them — and maintains the arena's public site, jetson-arena.com. Device order: Jetson Thor first, then Jetson AGX Orin on JetPack 7.2, then Jetson Orin Nano Super 8GB.

Status

Early. The benchmark domain surface is not built yet. What exists today is the agent-first CLI scaffold below, the mesh identity, the CI baseline, and the public site (jetson-arena.com, site-astro/) — there is no benchmark runner, device probe, or recipe store. The intended shape, and the questions still open, are written down in docs/scope.md. Read that before proposing a design.

Quickstart

uv sync
uv run pytest -n auto                 # run the test suite
uv run jetson-arena whoami            # identity from culture.yaml
uv run jetson-arena learn             # self-teaching prompt (add --json)
uv run teken cli doctor . --strict    # the agent-first rubric gate CI runs

CLI

Every command supports --json. Results go to stdout, errors/diagnostics to stderr (never mixed). Exit codes: 0 success, 1 user error, 2 environment error, 3+ reserved.

How it is put together

• An agent-first CLI cited from teken (afi-cli) — the runtime package has no third-party dependencies, and the contracts above are enforced in CI by a rubric gate that drives the built CLI as a black box.
• A mesh identity — culture.yaml (suffix + backend) and the matching resident prompt file (AGENTS.colleague.md, since this agent runs backend: colleague). CLAUDE.md is the parallel prompt for Claude Code.
• The guildmaster skill kit (14 skills) under .claude/skills/, vendored cite-don't-import. Provenance ledger: docs/skill-sources.md.
• A build + deploy baseline — pytest, lint, the agent-first rubric gate, and PyPI Trusted Publishing wired into GitHub Actions.

Contributing

Every PR bumps the version — docs-only ones included. CI enforces it, because a merge to main publishes to PyPI and a repeated version is a failed publish. See CLAUDE.md for the full conventions: the CLI contract, the vendored-skills rule, and the cicd PR lane.

License

Apache 2.0 — see LICENSE.